Total inference spend divided by tasks that produced an accepted result.
Total inference cost ÷ count of tasks whose output was accepted